While the atomic version might be better, at the moment the priority is
to just to get RLCT working reliably.

This is because we don't yet count the number of waiting threads,
instead just flagging where or not there are any waiters.

Maybe it would be better to simply stick to a mutex and condvar, as
libstd does. Optimization is for later anyway.

Otherwise it would be possible to do e.g.

`pthread_mutex_t mutex = PTHREAD_ONCE_INITIALIZER;```,
which would expand to
```pthread_mutex_t mutex = {0};```.